This is a high-fidelity dual-character LoRA trained for Wan 2.2, designed to showcase just how flexible and reliable multi-character generation can be when the training pipeline is done right.
The two characters were intentionally chosen for maximum contrast, making both identity separation and character consistency immediately obvious:
Character A: Goth woman in black fishnet and dark, high-contrast clothing
Character B: Handsome man in a clean, tailored business suit
Even in complex scenes, dynamic poses, or longer prompt chains, each character stays visually and stylistically distinct.
🎬 Why This LoRA Is Special
This isn’t just a “dual character test.”
It’s a demonstration of what you can do when you stop fighting identity bleed and start embracing multi-character storytelling:
Characters can interact naturally
Scene composition stays coherent
Clothing, facial features, and vibes stay locked to the correct character
You can build narrative continuity, not just single images
This opens the door to:
Story-driven image sequences
Multi-character scenes with emotional beats
Consistent recurring characters across generations
Wan 2.2 handles this extremely well when paired with the right training approach — and this LoRA is proof of that.
🎯 Intended Use
Multi-character storytelling
Dual-character scene generation
Testing identity consistency in Wan 2.2
Reference model for advanced LoRA training
Pushing beyond “one character per image” limitations
🔞 SFW / NSFW Notes
This LoRA can be used for both SFW and NSFW generations, depending on your prompts and base model.
Important clarification:
It was not specifically trained on NSFW content
Any explicit output comes from prompting and base model behavior, not the dataset
🛠️ Want to Create Your Own?
This LoRA is an example output of my dual-character training workflow.
If you want to create your own multi-character LoRAs — whether for storytelling, world-building, or more expressive scenes — the full workflow is available on LoboForge:
👉 https://www.loboforge.com/@LoboForge
The workflow covers:
Dataset structure for multiple characters
Caption separation techniques
Identity isolation strategies
Training settings for Wan 2.2
Common failure modes (and how to avoid them)
⚠️ Final Notes
This LoRA is designed to hold up under stress — complex prompts, dynamic scenes, and character interaction
If you’ve struggled with identity bleed before, this will immediately show what’s possible
Clean prompting + solid training = dramatically better results
Description
Prompt:
Ken is a handsome young man in a business suit.
Dakota is a pretty young goth girl with long black hair and fishnet.
Start with one or both of these and then say what they are doing.Â
Example:Â
Ken is a handsome young man in a business suit.
Dakota is a pretty young goth girl with long black hair and fishnet.
Ken is feeding Dakota strawberries. FAQ
Comments (17)
i havent tried this yet, but this concept is amazing. how do i learn to train them if i want to make such loras??? Great work by the way
I created a training workflow SPECIFICALLY to train this kind of LoRAs - its available on my Patreon for subscribers. Here https://www.patreon.com/posts/148254255
I did it by creating three datasets. Two sets of the characters alone, like you make a normal character lora, and one set of the two characters together. I had the together dataset be 50% the size of the single character datasets and set number of repeats to double that of the single character datasets.
For the together datasets I just pasted or edited the two characters together in GIMP making sure they have the same scale and it worked surprisingly well.
Make sure to separate the characters in your training prompts and use a low learning rate and you should be good.
Is there a A14B , High noise? , I see a 5B high noise and a A14B Low noise?, is that intended?
They are both from the same training session, i might have forgot to mark the second upload appropriately.
Hi LoboForge，I test this Lora at Wan2.2 14B，It can not work，so mabe it is a 5B Lora,right?
@LoboForge please correct it at least now, cuz I'm still confused
@honryindian its been corrected
@honryindian fixed
@LoboForge Thank you :)
is it possible with your method to achieve 3 consistent character
It is, though you would have to alter the training a bit - more steps, slightly adjusted learning speed and different types of images... I could probably create a workflow for this.. though it would be more complicated than the dual character workflow
Just what I need except that I would like it to work with any two characters. I have little need for these two pre-defined characters, nor would it be desirable to have to create a lora for each and every couple. Nevertheless, a good step in the right direction and I hope you can develop the idea.
Try the workflow - i actually just started with one image of each person. The workflow generated ALL the images i used to train this lora. The workflow is what you need, the lora is just an example of what you can produce with the workflow
@LoboForge In that case I'm downloading the workflow now... (ah, ok, paywall. Sorry but no). Thanks for your reply.
Have you tried two chars of the same gender? I imagine if they are different enough and trained/captioned well you may be able to get something working. I guess you'd need images of them together as well but I've never tried
I have - i created a lora of two girls as a test at first - i wasnt able to post it on civitai. https://www.patreon.com/posts/148702811
Here is the girls lora - its obviously SFW. I am working on a NSFW now and it will allow for man and woman, woman and woman, and man and man.